Modeling CBR Representation for Architectural Design ReusePaul Raduma, Lic. Tech., M.Arch., B.Arch.

Helsinki University of Technology, Espoo 02150 Finland, E-mail: Paul.Raduma@hut.fi

AbstractThe design of a good representation scheme is an impor-tant step towards facilitating the construction of case-based reasoners. The paper presents a design case repre-sentation scheme based on a conceptual CBR model forarchitectural design re-use, and discusses relevant lessonsfrom case representation in other CBR models. Themodel presented decomposes a house design, and usesrelative room positioning, co-ordinates and orientation torepresent the basic spaces within the house. The authorconsiders cases stored as memory of the design field.These cases are treated with a retrieval process (selectionand adaptation) from the memory. The originality of thisprocess is in the dynamic cycle of selection and adapta-tion. This paper uses the conceptual model to introduceand present a process of the dynamic cycle of selectionand adaptation of design cases.

1. IntroductionArchitects frequently use experiential knowledge in theproblem solving process. Case-Based Reasoning (CBR) based on the use of situational experiences in solving newproblems. There is growing interest in CBR and knowl-edge-based systems in the design professions. Applicationsrespired by this interest include ARCHIE (Domeshek andKolodner 1997) JANUS (Fisher, McCall, and Morch1989), SEED (Flemming et al. 1997), FABEL (Voss1997), and PRECEDENT (Oxman 1996). HouseCAD, conceptual CBR model discussed here, is part of an on-going research at Helsinki University of Technology. TheHouseCAD model uses relative room positioning, co-ordinates and orientation to represent the basic spaceswithin the house. It is based on snapshots from the designprocess, using the house’s plan view in its representation.The model assumes human-architect’s intervention in theadaptation process, based on the proposals from iterativeretrievals by the case based reasoner.

Design cases, like other cases in general have three ba-sic components: a situation or problem description, a solu-tion, and an outcome. The problem description records thestate of the world at the time the case was taking place, andif appropriate, what problem was to be solved at that time.

The solution records the solution to the problem specifiedin the problem description. The outcome records the re-suiting state of the world after the solution was carried out(Kolodner 1993). Cases with situation description and out-come can be used in evaluating new situations. This is es-pecially important in a continuous environment like archi-tectural design where it is not obvious where a case startsand where it ends. The design process of a building is con-tinuous, and changes are inevitable in all stages - frompreliminary to detailed design and construction phase. Asnapshot of a stage in the design process can be repre-sented as a case if it teaches a lesson for the future.

Case storage in CBR systems should reflect the con-ceptual view of what is represented in the case. In House-CAD, design-case representation for CBR indexing startswith extracting indexes from the bubble diagrams archi-tects employ in conceptual design. This information is laterused by the nearest-neighbor retrieval module, which com-pares the problem brief with design cases in the case mem-ory. The Case-based reasoner recommends to the humanarchitect the design cases suitable for further adaptationand storage.

In CBR, the case-base should be organized into a man-ageable structure that supports efficient search and retrievalmethods. A balance has to be found between storing meth-ods that preserve the richness of cases and their indexesand methods that simplify the access and retrieval of rele-vant cases. The methods are largely theoretical and hardlyapply in the commercially available CBR tools. Instead,the commercial CBR tools either store cases as simple andfiat file data-structures, or within conventional relationaldatabase structures and use indexes to reference the cases.The commercial CBR tools generally use two techniques:nearest neighbor retrieval and inductive retrieval. This isthe process of incorporating what is useful to retain fromthe new problem-solving episode into the existing knowl-edge. The learning from success or failure of the proposedsolution involves how to index the case for later retrievalfrom similar problems, and how to integrate the new casein the memory structure. Hence, the need to organize thecase-base into a manageable structure.

From: AAAI Technical Report SS-00-03. Compilation copyright © 2000, AAAI (www.aaai.org). All rights reserved.

2. Representing a Design Case

2.1 Bubble RepresentationArchitects love graphic representations, and often developtheir early design ideas using bubble diagrams. TheHouseCAD model proposes case-representation derivedfrom the bubble diagram. A bubble is drawn to representeach space; dark thick lines are drawn to indicate strongrelationships, while light thin lines represent weaker rela-tionship (Figure 1).

i~ ~;!::i:iiii;~iiiii!i~i!i~! ~!iii! {~{!i i!i~i!i~i!i!i{ i:i!i! i:i~{ i: i: i~ii!i~i!i!i~i!i~i!i!i!i~i~i!i!i j j!!!i~i!i~i!i!i!i~i!i!i~i~i!i~i~i~i:i~i!!! iii!i~i{ i: i: i~ !:!{i{i!i! i~ i:!j~!! i:i!i~i! i:i~i!i!i!

Figure 1. Bubble Representation of a Design Scenario

In HouseCAD, the bubble diagrams are defined by el-lipses whose centers are stored as 3D co-ordinates, andhave their attributes stored in a CAD program, and corre-spond to the center point of the respective space. The sce-nario shows a stage in the design process of the upper floorof a two-story house. The master bedroom is representedby B1, while B2, B3, and B4 represent the second, thirdand fourth bedrooms respectively. T1, T3 and PT representthe master bathroom, third bathroom and patio/terrace re-spectively.

In ARCHIE-II - a case-based design aid for the con-ceptual design of buildings - for example, a bubble editorassists the architect with conceptual design in three ways(Griffin and Domeshek 1996). First, to improve on a repre-sentational technique architects are already using by auto-mating editing and preserving an on-line record of theirwork. Second, to help the architects describe features of thespaces. Third, the bubble editor helps to form queries to thecase-based design aiding system. CBR systems use indexesto speed-up retrieval. A good representation scheme shouldreflect the conceptual view of what is represented in thecase. The CLAVIER system for autoclave loading (Henne-sey and Hinlde 1991, Hinlde and Toomey 1994) is one ofthe first commercially fielded CBR applications. It has itscase representation such that each layout case is described

in terms of the name of parts; tables on which the partswere placed; relative position of other parts; and produc-tion statistics such as start and finish times, pressure, andtemperature. In HouseCAD, the bubble diagram recordsone interpretation of the brief: the problem. The case li-brary stores design-cases: solutions that are representedpartly based on the bubble diagram.

2.2 Choosing the Design Cases

To determine the relative frequency of various spaces orrooms in home designs, an analysis was done on 612 con-temporary styles. These contemporary houses were takenout of an on-line case base of 5,005 from various design-ers. These had two stories or more, attached garages andhad floor area of between 92 and 278 m2. Out of these 46had floor areas of were between 255 and 275 m. Thisgroup was first studied to determine space frequency,which in turn would determine the 32 spaces and rooms tobe considered for indexing in the HouseCAD design-caserepresentation scheme (Raduma 1999).

2.3 Trimming the Case BaseThe selection criteria from the case-base of the 5,005 wasmodified such that only single story cases with 3 or 4 bed-rooms, attached garage, and with floor area of 135 to 185ms would be retrieved. The resultant was a case base of141, indexed 1001 to 1141. Sixteen of these cases wereremoved from the case base if, in the author’s opinion, theywere not good cases. The final case base used for the rep-resentation scheme had 125 cases. The cases were drawnfrom a total of 18 design companies. By limiting the designcases to single story cases with 3 or 4 bedrooms, attachedgarage, and with floor area of 135 to 185 ms, it makes itpossible to avoid noisy cases. It also eliminates cases thatwould otherwise make the design case base too largemanlmny enter into a case memory.

3. Decomposing a Design CaseBubble representation of a transformation of the brief isshown in Figure 2. The scenario shows the house dividedinto four quartiles - northeastern, southeastern, southwest-ern and northwestern quartiles. The plan north is for designcase representational and retrieval purposes only. Laterdiscussion shows how the best match for this transforma-tion compares with other case matches in the simulation(Figure 6).

/ \~, /...’"~. [ " .;:e..

..<. I\ Lr .... <~::::*C2 ......... ",\ J \ ~n~,

, .::. .~ i ~ ........., ,, p,$~.li~i \, ::: t i ":’::~P.~:’- ’ I". ~"’~t’rxP; ’t R~I ~ I ,:.> ’:::i~::..~ i:~,..~’~:" ~ ,’ ,- \

.............. ......... ........... CG .......i. ,’:~ % .::::> i ’t "~:: .... .,, i ~ "~,;("k~/;"::., , .... ::i:~.s:. J t] "::i:i:i~4~K..Z!,.,’ ~. " }

".=~=--. \ ~>.’~::" ~ :.>.:"’~ "T.x ’x ’....... ~.::.<.:.:, .~i ...... , ,.I/ \ \.,.~:;’/,,. <. ~. \-.,. .......

: B2 ’/A;’.,\. ~> J’x. ~ ..,.,~ ~,~s;’ik .~.: /,.,..-.,., -,=,,,

’ ~" I

Figure 2. Space Orientation and Location: Bnef 2

The case representation is an interpretation of the briefthat calls for a detached single-story, one-family house, fora retired couple. It should have three bedrooms, and an ex-tra room that may be used as an office. There is a spec-tacular view to the north. The living room, a coveredporch or deck, and one or two bedrooms should take ad-vantage of this view. Vehicular access to the house will befrom the southwest. It should accommodate a two-car ga-rage, and the lady of the house prefers it to be easily acces-sible to the kitchen and utilities area. Two separate walk-in closets are required in the master bedroom

An earlier model of HouseCAD divides the house intofour quartiles - northeastern, southeastern, southwestern ornorthwestern (Raduma 1999). It considers only one majororientation or view. A space has to be oriented to the north(N), south (S), east (E), west (W), oriented inward non-existent (X). The model represents each house plan sets of four representation schemes, transformation by re-flection of the original design-case. The transformation ofeach plan yields three additional versions of the original.The three additional cases may be necessary if orientationwith respect to the sun, wind, slope; view and other siteconstraint dictate, while other factors governing layout re-main constant. Depending on the transformational plan re-trieved, the living room of the same design-case originalmay be indexed to be in the NEQ, SEQ, SWQ, or NWQ,all of which are correct, and may apply in different sitecontexts. Transformation by reflection or rotation is notuncommon in site planning, and is made easy by CADprograms. A record of such transformation is important,and shows, after further adaptation how each unique designhas its own unique model. Each design process includesseveral design solutions that can also be represented ascases. As designers continually re-use and modify their

previous designs, it is important to model and analyze suchrecord (Raduma 1999). This increases the case memoryfrom which iterative solutions of design problems can beapplied.

ARCHIE uses specific space types based on courthouseexperiences. SEED relies on two central constructs - designunits and functional units in its indexing and retrieval. Inthe FABEL domain, a case is a layout, and its features areattribute-value pairs, or keywords that are associative. ACBR tool in the architectural design domain should aim todevelop a decision support system for architects. FABELfor example, explores various methods for combined rea-soning and mutual support of different knowledge types.HouseCAD uses major spaces and rooms in a house, andtheir orientation for indexing.

4. Case Descriptors

The HouseCAD represents its design cases by concentrat-ing on the solution. This bears some similarity with JULIA(Hinrichs 1992) and KRITIK (Goel 1989), which are design problem solvers. JULIA’s design task is in creatingmenu solutions. Most of the problems are too large, andhence are broken into parts. KRITIK’s problem domain isphysical devices (e.g. simple electrical circuits), and smaller than JULIA’s. Its functional representation showshow the pieces are connected to each other, and shows thegeneral effects of those connections.

~L~¢~;:le*lll>~’t¢’~l~’Ik~l’~’~ll’::,..:::.’:’’:::L ,::" :"’::’ ::: "’" :" ":: :’": ’ "...:’...i:!. "" ’~J~,~’$1~

........ i- ..................r-d;:z .......-~--~,: ...........................-r~

................ rrs¢~

::z:?;,’;:-:)~ lIT DH DKDIG ~ ~IC Ix

,.-.-:,’.=.:.-:~- , .......... ~ ..../ 2~d ~00r :-~,#’;’:lba’l.’’S l)Lo~ el~; B~lltu~O01~.]. Ills HOT ~J~/~PA

"-" ~’--%,."t...’:...:.~, era~ Ioa’el¢oo~ a ~: ~)Iloo~

::::::::::::::::::::::::::::::: :~: ~::::LI::::: :i:i::::~.::i:::~:i~:;4.-...;...;-,:-, ~..-.:~.:.e:-~, ~ .:. ̂-..:-.-:,-.:.,:~ ~ ,.;...J.......:., ;....i...:~ ~.....-.:.. ~.~-’:..-L J...:~,’.-:5¯11::_~6.’@:.~..:..:..;.:@Z.:..,.;~.Z. ’~.;.:.:..:..:..;.;.~..:..:..:..:~..’.... -~.~.i.~;..~i~..i....:.:,,..i..~.i.~.,-.~,.i.i.~..~..i..i.i. ~.,i..i..:.,.. :., .!..~

...................................................................................................................... i: 7 7lI~ ~~ ~’~an’~9"x/~ ~- ....... -2.

Figure 3. Design case representation

The representation scheme uses the 32 case descriptorsas discussed above. For simplicity, this scheme uses spacesand rooms on the main floor in its representation. Asshown in Figure 3, it is however, possible to use a moreelaborate 3-floor representation scheme (Raduma 1999).This scenario shows 3-D representation of the case de-3

scriptors, to show whether they represent spaces on thebasement (level 0) the main floor (level 1) or the upperfloor (level 2). This scenario is based on a study that used 3-metre by 3-metre grid. The model decomposes a housedesign, and uses relative room positioning, co-ordinatesand orientation to represent the basic spaces within thehouse.

The earlier HouseCAD design case representationmodel has two priority sets, each with 16-room or spacedescriptors (Figure 3). The first set of 16 contains the de-scriptors for the living room (LV), master-bedroom (B1),kitchen (KT), dinning room (DN), and main bathroom(T1). Other are the main stairs (S1), bedrooms 2 (B2) (B3), second bathroom (T2), family room (FR), walk-incloset (WI), loft/balcony (LB), Deck (DK), third bathroom(T3), Den/Library/Study/Office (DL), fourth bedroom and patio/terrace (PT). The model uses simplified 2D co-ordinate system, in addition to the three planes representingthe basement, first and second floors. This in effect is a 3Dmodel that uses relative room positioning to determine thelocation of the respective rooms within an eight by eightgrid. Master bedroom width determines the grid dimension- one unit on the X and Y-axes, while 1 in the Z-axis refersto level 1. Basement floor is level 0, while the upper flooris level 2 in the Z-axis.

5. Retrieval and MatchAssuming that the bubble diagram (Figure 1) is an inter-pretation of the brief - the problem to be solved - in termsof the required layout, the retrieval may yield a nearestneighbor such as the one shown (in Figure 4). In this hy-pothetical example, the retrieved case has some similarityin the positioning of master bedroom B1, bedrooms 2, 3and 4, bathrooms T1 and T2. In real life scenario, House-CAD may retrieve a design case from the case memory,which is not the nearest neighbor. CAD Layer use makes itpossible to compare this retrieved case with the first bubblediagram. The spatial relationship between the two bubblediagram representations is checked.

In terms of orientation, the case satisfies the desiredconditions in the master bedroom 1, bedrooms 2, 3, 4, sec-ond bathroom T2, and to some extent, the master bathroomT 1. One other aspect, that is, the distance between compa-rable rooms or spaces is computed. The distance betweenthe respective main stairs S 1, master bedroom B 1, and bed-rooms B2, B3, and B4 work out to be about 4.5 m, 3 m, 2.2m, 3.2 m and 5 m respectively. The human-architect con-siders that these dimensions are too big, probably removesthe design case from the case memory and another searchprocess is instituted.

After several searches from the existing case base, abetter design case is retrieved, whose bubble diagram rep-resentation is again compared with the original bubble rep-resentation of the brief (Figures 5). The distances are muchmore acceptable (about 0.9 m, 1.8 m, 0.8 m, 2.1 m, and 2.5m). Using weighted nearest-neighbor retrieval technique,the latter distances (2.1 m, 2.5m), that between respectivebedrooms 3 and 4, HouseCAD advises that the retrievedcase is acceptable for adaptation.

In the current HouseCAD representation scheme, eachdesign case is transformed by reflection and/or by rotation.This transformation yields an additional 7 design case rep-resentations. There are two alternatives: Each of the 125design cases with 8 design case representations and storedin the case-base, or a second alternative of representing thebrief with additional 7 design case representations. Thelatter has been used in the simulation. Hence, we have atotal of 8 design case representations namely the BRIEF,BRIEF 2, 3, 4, 5, 6, 7 and 8. The design Brief 7, for exam-ple, refers to the brief mirrored along X-axis with 90° rota-tion.

Each of these representations was matched with each ofthe 125 design cases in the case base. The HouseCAD rep-resentation scheme uses the 32 case descriptors discussedin step 4 above. An analysis of the 125 design cases re-vealed that the first 8 descriptors accounted for 985 out of2,106 instances (46.8%). The next 8 descriptors accountedfor 796 instances (37.8%) and the last 16 accounted for 325(15.4%). This information was used in weighed matching.Hence deign case 1,082 has a 23% match with the brief. Ithas 2 matches using the first 8 descriptors, 1 match withthe next 8, and one match with the last 16 case descriptors.

¯ i~.;::ii~ . ":ii .......

I .i ........i.i .......i.. :!..( ........ ~....!.:...i..~.i..~i...i....i...:.iN

illllllllll__ ::::::::::::::::::::::::::::::::::: ......~ ..........:::::::::::::::::::::::::::::::::::::: ......

z

Figure 5. Case Comparison

Figure 6. Case Retrieval 1122 (37% match)

The brief has a total of 18 case descriptors distributedas follows: 7 out of 8 in the first set; 7 out of 8 in the sec-ond set, and 3 out of 16 in the third set. Using the designcase representation of the brief, there was an averagematch of 10.9%. The best four retrievals had matches of39%, 34%, 28% and 28%. There was no match at all in 18out of the 125 cases. The design Brief 2, a transformationof the brief - mirrored along Y-axis had an average matchof 12.7% - better than the original brief. The best four re-trievals had matches of 37%, 37%, 36% and 35% (Figure6, Table 1).

There was no match at all in 16 out of the 125 cases.Brief 3, a transformation of the brief mirrored along X-axishad an average match of 6%. Brief 4, a transformation ofthe brief mirrored along Y-axis & with 90° rotation had anaverage match of 5.6%. The best four retrievals hadmatches of 18%, 16%, 16% and 16%. There was no matchat all in 39 out of the 125 cases Brief 5, a transformation ofthe brief mirrored along X-axis had an average match of5.9%. Brief 6, a transformation of the brief with 90° rota-tion had an average match of 4.3%. Brief 7, a transforma-tion of the brief mirrored along X-axis with 90° rotationhad an average match of 6.1%. The best four retrievals hadmatches of 23%, 23%, 17% and 16%. There was no matchat all in 42 out of the 125 cases. Brief 8, a transformationof the brief with 180° rotation had an average match of3.4%.

BRIEF 1 BRIEF 2max. match total max. match total

18 18weight 100 47 100CASE case match % CASE case match %1065 3 4 0 39 1102 2 5 0 371074 3 3 0 34 1122 2 5 0 371026 2 3 0 28 1123 1 4 2 361047 2 3 0 28 1037 3 2 1 351052 1 4 0 26 1062 3 3 0 341007 3 1 0 :25 1063 3 3 0 341008 3 1 0 125 1075 3 3 0 341009 3 1 0 125 1106 1 4 1 311086 3 1 0 125 1112 1 4 1 311082 2 1 1 23 1107 1 5 0 3O1006 2 2 0 23 1051 3 2 0 3O1121 2 2 0 23 1081 2 3 0 281028 1 2 1 21 1104 2 3 0 281010 1 3 0 21 1068 1 3 1 261032 1 3 0 21 1101 1 3 1 26

Table 1. Best 15 matches in Brief I and Brief2

The result of the simulation, especially of the Brief andBrief 2 is significant. It should be noted that the case basewas restricted to singe story houses with 3 or four bed-rooms and an attached garage. Each of the 32 case de-scriptors can be represented to be in any of the four quarti-les (NEQ, SEQ, SWQ and NWQ). ff represented in each the four quartiles, the case descriptor may be represented toopen to one of the nine orientations: the north, northeast,east, southeast, south, southwest, west, northwest, or in-ward. This implies that there are 4*9 alternatives in repre-senting each descriptors, or 36 possibilities. The brief had18 spaces, each of which has 36 orientation alternatives.

5

The number of combinations and permutations is huge. Onaverage, the 125 design cases had a little less than 18 casedescriptors, but some had as little as 14 (case 1022). achieve an exact match, the correct design case shouldhave 18 case descriptors in a specific orientation. Thechance for getting an exact match in any of the 125 designcases is close to zero. The success of our case base is dueto the fact that only cases relevant to the narrow domain(singe story houses with 3 or four bedrooms and an at-tached garage) was used. Further the cases were repre-sented such that the entrance lobby (and front porch) wereoriented to the south, thus limiting any random location ofthe lobby (LB) and in some cases, the front porch (C1) the south.

5. Adaptation QuestionAdaptation is widely accepted as the biggest challenge inthe CBR process. It is yet to be seen if case-based reason-ers in the design domain will develop as autonomous sys-tems that carry, out retrieval, adaptation and evaluation pro-cess by themselves. We have a lot to learn from, for exam-ple, CHEF, JULIA, PLEXUS, CASEY, PROTOS andHYPO, case-based reasoners considered by some to beautonomous systems (Leake, 1996). Human-machine sys-tem that works along with architects to solve design prob-lems or interpret design scenarios is proposed in theHouseCAD model. CLAVIER, considered successful inthe design domain by most CBR scientists, was designed toadapt autoclave layouts, but the engineers on the grounddid not like the adaptation facility. They opted to adapt thelayouts themselves, using CLAVIER to check that theirhuman adaptation will not repeat a failed autoclave layout.

The adapted design-case, from its case representationor indexing, is compared with other indexes in the casememory. In one scenario, one design case compared veryclosely with the original bubble diagram - the brief (Figure1), especially the relative position and orientation of masterbedroom B1. It also compared very well with the adjacentbedroom B3, bedroom B4 with its "western" orientation,position of the patio/terrace PT and the position of themaster bathroom T1. The human-architect may accept thisas the final retrieval from which minor and final adaptationwill be made. The architect may alternatively retrieve an-other case, compare it with the original bubble diagram,transform it, and adapt it. In order to satisfy the constraintsset out earlier, the architect converts the master bedroomB1 to B2, then enlarges B2 to the size of B1, and convertsthe master bathroom T1 to patio/terrace PT. The resultantgoes through transformation by reflection to yield a total offour new designs. This is stored as part of the case-base(Figures 7, 8).

The adaptation method in HouseCAD is not autono-mous, but based on the human architect’s intervention. Themethod is a combination of specialized search and case-based substitution. Both auxiliary knowledge structuresand case memory are queried, and specialized search heu-ristics are used to guide the memory search. Case-basedsubstitution uses other cases to suggest substitutions as inCLAVIER and JULIA.

6. DiscussionArchitectural design re-use is one avenue of CBR researchthat can be of direct benefit in housing design for largepopulation groups. The traditional white table approach tolow-income housing design cannot yield user-specific de-signs if there is limited time and financial resources com-mitted to architectural design. The white table approach in

6

housing design for such large population groups may resultin duplicating inappropriate prototypes in hundreds orthousands, not suited to varying user profiles. The designof housing has to be seen from the design of the individualunit and as part of a larger urban planning project. It istherefore useful to compare lessons from case-based plan-ning. In the design and planning of housing, CBR as aknowledge-based system is feasible as a knowledge assis-tant, and an intelligent consultant (Yeh and Shi 1999). case-based planning (CBP), previously generated plans arestored as cases in memory and can be reused to solvesimilar planning problems in the future. CBP can save con-siderable time over planning from scratch (generativeplanning), thus offering a potential (heuristic) mechanismfor handling intractable problems. One drawback of CBPsystems has been the need for a highly structured memorythat requires significant domain engineering and complexmemory indexing schemes to enable efficient case re-trieval. This problem can be alleviated, like in CAPER(Hendler 1995), using a CBP system that is based on massively frame-based AI language that can do extremelyfast retrieval of complex cases from a large, non-indexedmemory.

The performance of Case-Based Reasoners, like otherintelligent systems, depends on factors including inherentdesign decisions and domain characteristics. In the archi-tectural design domain, breaking the task in to small man-ageable pieces best solves the problem. This does not pre-clude looking at the big picture- how the individual partsrelate to each other and with the outside world. TheHouseCAD model proposes a design case representationmodel in the domain of residential architecture based onthe decomposition of a house design, and the use of rela-tive room positioning, co-ordinates and orientation to rep-resent the basic spaces within the house.

7. References

¯ Domeshek, E., and Kolodner, J. 1997. The Designer’sMuse. In lssues and Applications of Case-Based Rea-soning in Design, M. L. Maher and Pu, P. eds.Hillsdale, N.J: Lawrence Erlbaum Associates: 11-38.

¯ Fisher, G., McCall, R., and Morch, A. 1989. DesignEnvironments for Constructive and ArgumentativeDesign. In, Proceedings, Human Factors in Comput-ing Systems Conference, New York: ACM Press: 269-275.

¯ Flemming, U., Aygen, Z., Coyne, R., and Snyder J.,1997. Case-Based Design in a Software Environmentthat supports the Early Phases in Building Design. Inlssues and Applications of Case-Based Reasoning inDesign: 61-86.

¯ Griffith A.L. and Domeshek, E. 1996. IndexingEvaluations of Buildings to Aid Conceptual Design. InLeake, D. B., ed. Case-Based Reasoning. Experiences,Lessons & Future Directions Cambridge, MA: AAAIPress/MIT Press

¯ Goel, A. 1989. Integration of case-based reasoning andmodel-based reasoning for adaptive problem solving.Ph.D. dissertation, Dept. of Computer and InformationScience, The Ohio State University.

¯ Hendler, J. A. 1995. Evaluating the CAPER PlanningSystem, Proceedings, Israeli Symposium on ArtificialIntelligence, Jerusalem, Israel, Jan 1995.

¯ Hennesey, D. and Hinkle, D. 1991. Initial Results fromClavier: A case-based autoclave loading assistant. InProceedings of the DARPA Case-Based ReasoningWorkshop, edited by R. Bareiss. San Francisco:Morgan Kauffman Publishing.

¯ Hinkle, D., & Toomey, C. 1994. Clavier: ApplyingCase-Based Reasoning to Composite Part Fabrication.In Innovative Applications of Artificial Intelligence 6,54-62. Cambridge, MA: AAAI Press/MIT Press.

¯ Hinrichs, T.R. 1992. Problem Solving in open worlds.A Case Study in Design. Northvale, NJ: Erlbaum.

¯ Kolodner, J. 1993. Case-Based Reasoning. San Mateo,CA: Morgan Kanfmann.

¯ Leake, D. B. 1996. CBR in Context: The Present andFuture. In Leake, D. B. ed.. Case-Based Reasoning.Experiences, Lessons & Future Directions Cambridge,MA: AAAI Press/MIT Press

¯ Oxman, R. E. 1996. Case-based design support: sup-porting architectural composition through precedentlibraries. Journal of Architectural and Planning Re-search 13, 242-255.

¯ Raduma, P. 1999. Designing a CBR RepresentationScheme for Architectural Design Reuse. Arpakannus1/99, Newsletter of the Finnish Artificial IntelligenceSociety, Special Issue on the IJCAI’99 affiliate eventNetworks’99, Espoo, Finland August 8-19, 1999.

¯ Voss, A. 1997. Case Design Specialists in Fabel. InMaher M.L & Pu, P. eds. lssues and Applications ofCase-Based Reasoning in Design, Lawrence Erlbaum

¯ Yeh, A. G. O., and Sift, X. 1999. Applying case-basedreasoning to urban planning. A new planning-supportsystem tool. Environment and Planning B: Planningand Design 1999, volume 26 pages 101-115.